Legal claims defining the scope of protection. Each claim is shown in both the original legal language and a plain English translation.
1. A method comprising: producing, by one or more processors, a processed image based on a user-captured image by processing the user-captured image to focus on at least one object of interest depicted in the user-captured image, the object of interest comprising an article of clothing being worn by a human individual; training, by the one or more processors, a first image generator by processing the processed image using a first generative adversarial network, the processed image depicting the at least one object of interest, the first generative adversarial network comprising the first image generator, a first discriminator, and a second discriminator, and the training the first image generator by the first generative adversarial network comprises: generating, by the first image generator, a first generated image based on the processed image; determining, by the first discriminator, whether the first generated image is a real image based on a reference image; determining, by the second discriminator, whether the processed image and the first generated image are associated; and adjusting the first image generator based on a result of the determining by the first discriminator and a result of the determining by the second discriminator; combining, by the one or more processors, the processed image and the first generated image to produce a combined image; and training, by the one or more processors, a second image generator by processing the processed image using a second generative adversarial network, the second generative adversarial network comprising the second image generator and a third discriminator, and the training the second image generator by the second generative adversarial network comprises: generating, by the second image generator, a second generated image based on the combined image; determining, by the third discriminator, whether the second generated image is associated with the processed image; and adjusting the second image generator based on a result of the determining by the third discriminator.
2. The method of claim 1 , wherein the producing the processed image based on the user-captured image comprises: performing a person segmentation process on the user-captured image to produce a segmented image; disposing at least one segment of the segmented image on a white background to produce an intermediate image; performing a human pose estimation process on the intermediate image to identify a plurality of points of interest in the intermediate image; and cropping the intermediate image based on the plurality of points of interest to produce the processed image.
This invention relates to image processing techniques for enhancing user-captured images, particularly for isolating and refining human subjects within an image. The method addresses the challenge of extracting a person from a background while preserving key features for further analysis or display. The process begins by performing person segmentation on a user-captured image to separate the subject from the background, producing a segmented image. This segmented image is then placed on a white background to create an intermediate image, which simplifies subsequent processing. A human pose estimation process is applied to the intermediate image to identify key points of interest, such as joints or facial landmarks. These points are used to determine optimal cropping boundaries, ensuring the subject is framed accurately while minimizing unnecessary background. The final processed image is generated by cropping the intermediate image based on these identified points, resulting in a clean, focused representation of the subject. This technique is useful in applications like portrait enhancement, virtual try-on systems, or augmented reality, where precise subject isolation and framing are critical.
3. The method of claim 2 , wherein the article of clothing comprises a clothing top being worn by the human individual, and the plurality of points of interest correspond to a torso portion of the human individual.
This invention relates to a method for analyzing human body movements using wearable sensors, specifically focusing on tracking the torso region of an individual wearing a clothing top. The method involves detecting and monitoring multiple points of interest on the torso to capture movement data. These points are strategically placed to ensure accurate tracking of the torso's motion, which can be used for applications such as posture analysis, medical monitoring, or fitness tracking. The sensors embedded in the clothing top collect data from these points, allowing for real-time or post-processing analysis of the individual's movements. The system may include additional components, such as data processing units or communication modules, to transmit and analyze the collected data. The method ensures precise tracking by focusing on the torso, which is a critical area for understanding overall body mechanics and health-related metrics. The invention addresses the challenge of accurately capturing torso movements in a non-invasive and wearable form, improving upon traditional methods that may rely on bulky or less precise equipment.
4. The method of claim 1 , wherein the first image generator comprises a first encoder-decoder neural network and the second image generator comprises a second encoder-decoder neural network.
This invention relates to image generation using neural networks, specifically addressing the challenge of generating high-quality images from input data. The method involves two distinct image generators, each implemented as an encoder-decoder neural network. The first encoder-decoder neural network processes input data to produce a first set of image features, while the second encoder-decoder neural network independently processes the same or related input data to generate a second set of image features. These networks operate in parallel or sequentially to enhance image quality, resolution, or diversity. The encoder-decoder architecture allows each network to compress input data into a latent representation and then reconstruct it into an output image, with the two networks potentially specializing in different aspects of image generation, such as texture or structure. The method may be applied in applications like image synthesis, style transfer, or super-resolution, where multiple neural networks collaborate to improve the final output. The use of separate encoder-decoder networks enables flexibility in designing specialized architectures for different image generation tasks.
5. The method of claim 1 , wherein the training the second image generator by processing the processed image using the second generative adversarial network further comprises: calculating, by a L-2 loss function, a difference between the second generated image and the reference image, the adjusting the second image generator further based on the difference.
This invention relates to image generation using generative adversarial networks (GANs), specifically improving the quality of generated images by refining a second image generator with a reference image. The problem addressed is the difficulty in producing high-fidelity images from a GAN, particularly when the initial generator lacks sufficient detail or accuracy. The method involves training a second image generator using a second GAN, where the generator processes an input image to produce a second generated image. To enhance accuracy, the method calculates the difference between the second generated image and a reference image using an L-2 loss function. The second image generator is then adjusted based on this difference, refining its output to better match the reference. This refinement step ensures the generated image retains high fidelity to the desired reference, improving overall image quality. The process may also involve preprocessing the input image before generation, such as applying transformations or noise reduction, to further enhance the input quality before the second generator operates. The L-2 loss function provides a quantitative measure of discrepancy, allowing precise adjustments to the generator's parameters. This approach is particularly useful in applications requiring high-precision image synthesis, such as medical imaging, artistic rendering, or data augmentation for machine learning.
6. The method of claim 1 , further comprising: using, by the one or more processors, at least one of the first image generator or the second image generator to perform an image-based search for one or more images in an image database to produce a set of search result images, the image-based search being performed based on an input image.
This invention relates to image processing and search technologies, specifically addressing the challenge of efficiently retrieving relevant images from a database based on an input image. The method involves generating at least two distinct images from an input image using separate image generators, each applying different transformations or enhancements to the input image. These generated images are then used to perform an image-based search across an image database. The search is designed to identify and return a set of search result images that closely match the input image, leveraging the variations produced by the different image generators to improve search accuracy and coverage. The system may employ machine learning or pattern recognition techniques to analyze and compare visual features between the input image and database images, ensuring robust and context-aware retrieval. This approach enhances the ability to find visually similar or semantically related images, addressing limitations in traditional search methods that rely on single-image queries or text-based metadata. The invention is particularly useful in applications such as e-commerce, digital asset management, and content recommendation systems where precise image matching is critical.
7. The method of claim 6 , wherein the using the at least one of the first image generator or the second image generator to perform the image-based search for the one or more images in the image database to produce the set of search result images comprises: accessing the input image provided by a client device; processing the input image using the first image generator and the second image generator to generate an output image; and performing the image-based search using the output image.
This invention relates to image-based search systems, specifically improving the accuracy and efficiency of searching for similar images in a database. The problem addressed is the challenge of retrieving relevant images when input queries may vary in quality, context, or representation, leading to poor search results. The solution involves using multiple image generators to enhance the input image before performing the search. The method processes an input image provided by a client device using at least one of two image generators. The first image generator may apply transformations such as style transfer, noise reduction, or resolution enhancement to improve the input image's features. The second image generator may generate alternative representations of the input image, such as different artistic styles or synthetic views, to capture diverse aspects of the content. The processed output image, whether enhanced or alternative, is then used to search an image database, producing a set of search result images that are more relevant to the original input. By leveraging multiple generators, the system compensates for variations in input quality and context, improving the robustness of image-based search. This approach is particularly useful in applications like e-commerce, social media, or medical imaging, where accurate image retrieval is critical. The method ensures that the search results align closely with the user's intent, even when the input image is suboptimal.
8. The method of claim 6 , further comprising: identifying, by the one or more processors, a set of item records, in an item database of an online marketplace system, based on the set of item records being associated with the set of search result images; and causing, by the one or more processors, a set of items associated with the set of item records to be presented at a client device as one or more items suggested for purchase through the online marketplace system.
This invention relates to online marketplace systems that enhance search functionality by suggesting items for purchase based on visual search results. The problem addressed is the difficulty users face in finding relevant items when searching with images, as traditional text-based search may not capture the full context or visual attributes of desired products. The system processes a set of search result images obtained from an initial visual search query. These images are analyzed to identify a set of item records in an item database of the online marketplace. The item records are selected based on their association with the search result images, meaning they share visual or contextual similarities. The system then presents a set of items corresponding to these item records at a client device, suggesting them as potential purchases. This helps users discover relevant products that match the visual characteristics of their search query, improving the shopping experience by reducing the need for manual filtering or additional searches. The method leverages image recognition and database matching to bridge the gap between visual search inputs and product discovery, making it easier for users to find and purchase items that align with their visual preferences. The approach enhances the efficiency of online shopping by automating the suggestion of visually similar items, thereby increasing user engagement and potential sales.
9. The method of claim 6 , wherein the input image is provided for a listing of an item for sale on an online marketplace system, the method further comprising: causing, by the one or more processors, the set of search result images to be provided to a client device as one or more suggested images for use with the listing.
This invention relates to image processing and online marketplace systems, specifically improving the quality and relevance of images used in product listings. The problem addressed is the difficulty for sellers to select visually appealing and relevant images when creating listings, which can impact buyer engagement and sales. The method involves analyzing an input image, such as a product photo, to identify visual features and compare them against a database of search result images. The comparison is based on visual similarity, such as color, texture, composition, or object recognition, to determine a set of similar images. These similar images are then provided as suggested images to the seller, helping them choose better visuals for their listing. The system may also rank the suggested images based on relevance or popularity to further assist the seller. The method may include preprocessing the input image to enhance its features before comparison, ensuring accurate matching. The search result images can be sourced from existing listings, user-uploaded content, or curated databases. The suggestions are dynamically generated and displayed to the seller in real-time, allowing for quick adjustments to the listing. This approach improves the visual appeal of listings, potentially increasing buyer interest and conversion rates.
10. The method of claim 1 , further comprising: accessing, by the one or more processors, an input image provided by a client device; processing, by the one or more processors, the input image using the first image generator to produce a first intermediate image; combining, by the one or more processors, the input image with the intermediate image to produce a second intermediate image; and processing, by the one or more processors, the second intermediate image using the second image generator to produce a generated output image.
This invention relates to image processing systems that enhance or modify input images using generative models. The technology addresses the challenge of producing high-quality output images from low-quality or incomplete input images by leveraging multiple generative models in sequence. The method involves accessing an input image from a client device and processing it using a first image generator to produce a first intermediate image. This intermediate image is then combined with the original input image to create a second intermediate image. The second intermediate image is further processed using a second image generator to produce the final output image. The sequential use of two generative models allows for iterative refinement, where the first model captures initial enhancements or modifications, and the second model refines those results to produce a more polished output. This approach is particularly useful in applications such as image restoration, style transfer, or super-resolution, where multiple stages of processing can improve the quality of the final result. The system may be implemented in software running on one or more processors, with the ability to handle input images provided by external devices. The method ensures that the original input image contributes to the final output, preventing complete over-generation by blending intermediate results.
11. The method of claim 1 , wherein the combining the processed image and the first generated image to produce the combined image comprises concatenating the processed image and the first generated image to produce the combined image.
This invention relates to image processing techniques, specifically methods for combining processed images with generated images to produce a final combined image. The problem addressed is the need for efficient and accurate methods to merge different image representations while preserving relevant features from both inputs. The method involves processing an input image to extract or modify specific features, such as edges, textures, or other visual elements. A first generated image is created, which may be a synthetic or modified version of the input image, such as a stylized, enhanced, or reconstructed image. The processed image and the first generated image are then combined by concatenating them, meaning their respective data structures or feature representations are joined together in a structured manner. This concatenation may involve aligning corresponding features or merging pixel-level or feature-level data to produce a combined image that retains contributions from both inputs. The concatenation step ensures that the combined image integrates the processed image's modifications with the generated image's enhancements, resulting in a final output that leverages the strengths of both. This approach is useful in applications like image restoration, style transfer, or super-resolution, where combining different image representations can improve visual quality or accuracy. The method may be implemented using neural networks or other computational techniques to handle the image processing and concatenation steps efficiently.
12. A system comprising: one or more processors; and a computer-readable medium having instructions stored there on, which, when executed by the one or more processors, cause the system to perform operations comprising: producing a processed image based on a user-captured image by processing the user-captured image to focus on at least one object of interest depicted in the user-captured image, the object of interest comprising an article of clothing being worn by a human individual; training a first image generator by processing the processed image using a first generative adversarial network, the processed image depicting the at least one object of interest, the first generative adversarial network comprising the first image generator, a first discriminator, and a second discriminator, and the training the first image generator by the first generative adversarial network comprises: generating, by the first image generator, a first generated image based on the processed image; determining, by the first discriminator, whether the first generated image is a real image based on to a reference image; determining, by the second discriminator, whether the processed image and the first generated image are associated; and adjusting the first image generator based on a result of the determining by the first discriminator and a result of the determining by the second discriminator; combining the processed image and the first generated image to produce a combined image; and training a second image generator by processing the processed image using a second generative adversarial network, the second generative adversarial network comprising the second image generator and a third discriminator, and the training the second image generator by the second generative adversarial network comprises: generating, by the second image generator, a second generated image based on the combined image; determining, by the third discriminator, whether the second generated image is associated with the processed image; and adjusting the second image generator based on a result of the determining by the third discriminator.
The system enhances images of clothing worn by individuals to improve visual quality and consistency. The technology addresses challenges in capturing clear, well-focused images of clothing, particularly in real-world scenarios where lighting, background, or motion may degrade image quality. The system processes a user-captured image to focus on the clothing item, generating a refined version of the image. A first generative adversarial network (GAN) is trained using this processed image, where a generator creates a modified image, and two discriminators evaluate its realism and association with the original. The first discriminator compares the generated image to a reference, while the second checks if the generated and processed images are related. The generator is adjusted based on these evaluations. The processed and generated images are then combined, and a second GAN further refines the output. This second GAN includes a generator and a discriminator that assess whether the new generated image aligns with the original processed image. The system iteratively improves image quality, ensuring the final output accurately represents the clothing while maintaining natural appearance. This approach is useful for applications like e-commerce, virtual try-ons, or fashion design, where high-quality, consistent clothing images are essential.
13. The system of claim 12 , wherein the instructions further cause the system to perform the operations of: using at least one of the first image generator or the second image generator to perform an image-based search for one or more images in an image database to produce a set of search result images, the image-based search being performed based on an input image.
This invention relates to image processing systems that generate images and perform image-based searches. The system includes a first image generator and a second image generator, each capable of producing images based on input data. The system also includes a memory storing instructions that, when executed, enable the system to perform operations such as generating images using the first and second image generators. The system further includes a processor configured to execute the instructions to perform these operations. The invention addresses the challenge of efficiently generating and searching images in a database, particularly when dealing with large datasets or complex image queries. The system enhances image retrieval by leveraging multiple image generators to improve search accuracy and relevance. The image-based search is performed using an input image to query an image database, producing a set of search result images that match or are similar to the input image. This allows users to quickly find visually similar images within a database, improving efficiency in applications such as content management, e-commerce, and digital media. The system's ability to utilize multiple image generators ensures robust performance, even when dealing with diverse or high-dimensional image data.
14. The system of claim 13 , wherein the instructions further cause the system to perform the operations of: identifying a set of item records, in an item database of an online marketplace system, based on the set of item records being associated with the set of search result images; and causing a set of items associated with the set of item records to be presented at a client device as one or more items suggested for purchase through the online marketplace system.
This invention relates to an online marketplace system that enhances product discovery by suggesting items for purchase based on visual search results. The system addresses the challenge of helping users find relevant products when they lack specific search terms or prefer browsing visually. The system processes a set of search result images, such as those generated from a visual search query, and identifies item records in an item database that are associated with these images. These item records correspond to products available for purchase. The system then presents these products to a user at a client device as suggested items, facilitating easier discovery and purchase. The underlying technology likely involves image recognition, database querying, and user interface rendering to bridge visual search results with product listings. This approach improves user experience by reducing reliance on textual search and leveraging visual cues to recommend relevant items. The system may integrate with existing online marketplace features, such as search engines or recommendation algorithms, to provide a seamless shopping experience.
15. The system of claim 13 , wherein the using the at least one of the first image generator or the second image generator to perform the image-based search for the one or more images in the image database to produce the set of search result images comprises: accessing the input image provided by a client device; processing the input image using the first image generator and the second image generator to generate an output image; and performing the image-based search using the output image.
This invention relates to an image-based search system that enhances search accuracy by leveraging multiple image generators. The system addresses the challenge of retrieving relevant images from a database when input images may lack sufficient detail or clarity for precise matching. The system includes a first image generator and a second image generator, each capable of transforming an input image into different output images optimized for search purposes. The system processes an input image provided by a client device using both generators to produce distinct output images. These output images are then used to perform an image-based search across an image database, yielding a set of search result images. The dual-generation approach improves search robustness by compensating for limitations in either generator, ensuring more accurate and comprehensive results. The system may also include a user interface for displaying search results and a database for storing images and associated metadata. The invention is particularly useful in applications requiring high-precision image retrieval, such as e-commerce, medical imaging, or augmented reality.
16. The system of claim 13 , wherein the input image is provided for a listing of an item for sale on an online marketplace system, the instructions further causing the system to perform the operations of: causing the set of search results to be provided to a client device as one or more suggested images for use with the listing.
This invention relates to an image search system for online marketplaces, specifically addressing the challenge of helping sellers find relevant images for their product listings. The system processes an input image, such as a product photo, and generates a set of search results comprising similar or related images. These results are then presented to the seller as suggested images for their online marketplace listing, improving the visual appeal and accuracy of the product representation. The system may use image recognition, machine learning, or other computational techniques to identify visually similar or contextually relevant images from a database. The search results are dynamically provided to the seller's client device, allowing them to select and incorporate the suggested images into their listing. This enhances the efficiency of the listing creation process and ensures that the images used are appropriate for the product being sold. The system may also include additional features such as filtering, ranking, or categorizing the search results based on relevance, popularity, or other criteria to further refine the suggestions.
17. The system of claim 12 , wherein the training the second image generator by processing the processed image using the second generative adversarial network further comprises: calculating, by a L-2 loss function, a difference between the second generated image and the reference image, the adjusting the second image generator further based on the difference.
This invention relates to image processing systems that use generative adversarial networks (GANs) to improve image quality. The system addresses the challenge of generating high-fidelity images by refining an initial image through iterative training of multiple GANs. The system includes a first image generator that produces an initial image from input data, which is then processed by a second image generator to create a refined output. The second generator is trained using a reference image, with adjustments made based on a calculated difference between the generated image and the reference. Specifically, an L-2 loss function measures the discrepancy between the second generated image and the reference image, and the generator is further adjusted based on this difference to enhance accuracy and realism. The system leverages adversarial training to iteratively improve image quality, ensuring the final output closely matches the desired reference. This approach is particularly useful in applications requiring high-quality image synthesis, such as medical imaging, computer vision, and augmented reality.
18. The system of claim 12 , wherein the instructions further cause the system to perform the operations of: accessing an input image provided by a client device; processing the input image using the first image generator to produce a first intermediate image; combining the input image with the intermediate image to produce a second intermediate image; and processing the second intermediate image using the second image generator to produce a generated output image.
This invention relates to an image processing system that enhances or modifies input images using a multi-stage generative model. The system addresses the challenge of improving image quality, style transfer, or other transformations while maintaining computational efficiency and preserving key features of the original input. The system includes a first image generator and a second image generator, each trained to perform distinct image processing tasks. The first generator processes an input image to produce an intermediate image, which is then combined with the original input to create a second intermediate image. This second intermediate image is further processed by the second generator to produce a final output image. The combination of the input and intermediate images ensures that the original image's structure is retained while applying the desired transformations. The system is designed to operate in real-time or near real-time, making it suitable for applications such as photo editing, virtual try-on, or augmented reality. The use of two generators allows for more complex and refined image modifications compared to single-stage systems. The system can be deployed on client devices or cloud-based servers, depending on the computational requirements.
19. The system of claim 12 , wherein the combining the processed image and the first generated image to produce the combined image comprises concatenating the user-provided image and the first generated image to produce the combined image.
This invention relates to image processing systems that combine user-provided images with generated images to produce a final output. The problem addressed is the need for an efficient and effective way to merge these images while preserving their respective features. The system processes an input image, such as a user-provided image, and generates a first image using a generative model. The key innovation is in the method of combining these images. Specifically, the processed user-provided image and the first generated image are concatenated to produce a combined image. This concatenation ensures that the spatial and feature information from both images is retained in the final output. The system may also include additional steps such as preprocessing the input image, generating multiple images, and refining the combined image through further processing stages. The concatenation method is particularly useful in applications like image inpainting, style transfer, or super-resolution, where maintaining coherence between the original and generated content is critical. The approach improves upon traditional blending techniques by avoiding loss of detail or artifacts that can occur with other merging methods. The system is designed to work with various types of generative models, including neural networks, and can be adapted for different image processing tasks.
20. A non-transitory computer-readable storage medium comprising instructions that, when executed by at least one processor of a machine, cause the machine to perform operations comprising: producing a processed image based on a user-captured image by processing the user-captured image to focus on at least one object of interest depicted in the user-captured image, the object of interest comprising an article of clothing being worn by a human individual; training a first image generator by processing the processed image using a first generative adversarial network, the processed image depicting the at least one object of interest, the first generative adversarial network comprising the first image generator, a first discriminator, and a second discriminator, and the training the first image generator by the first generative adversarial network comprises: generating, by the first image generator, a first generated image based on the processed image; determining, by the first discriminator, whether the first generated image is a real image based on to a reference image; determining, by the second discriminator, whether the processed image and the first generated image are associated; and adjusting the first image generator based on a result of the determining by the first discriminator and a result of the determining by the second discriminator; combining the processed image and the first generated image to produce a combined image; and training a second image generator by processing the processed image using a second generative adversarial network, the second generative adversarial network comprising the second image generator and a third discriminator, and the training the second image generator by the second generative adversarial network comprises: generating, by the second image generator, a second generated image based on the combined image; determining, by the third discriminator, whether the second generated image is associated with the processed image; and adjusting the second image generator based on a result of the determining by the third discriminator.
This invention relates to image processing and generative adversarial networks (GANs) for enhancing and modifying images of clothing worn by individuals. The technology addresses the challenge of improving image quality and generating realistic variations of clothing items in user-captured photos, which often suffer from poor focus, lighting, or composition. The system processes a user-captured image to focus on an article of clothing worn by a person. A first generative adversarial network (GAN) is trained using the processed image, where the GAN includes a generator and two discriminators. The generator creates a modified version of the clothing item, while the first discriminator evaluates its realism against a reference image, and the second discriminator checks if the generated image matches the original processed image. The generator is adjusted based on feedback from both discriminators. The processed image and the generated image are then combined, and a second GAN is trained using this combined image. The second GAN includes another generator and a discriminator, which assesses whether the newly generated image aligns with the original processed image. The second generator is refined based on this evaluation. This approach enables high-quality image enhancement and realistic clothing modifications, useful for applications like virtual try-on, fashion design, and e-commerce. The dual-GAN architecture ensures both realism and consistency with the original image.
Unknown
February 4, 2020
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.